As Bernstein explains in [SWN] just about everything I need to know in order to plan for retirement isn't just unknown, it's unknowable. Geometric average stock returns? Distribution pattern for bond returns? Correlations between stocks and bonds? Nobody knows and it looks like no one actually can know. My solution to this conundrum is relatively straightforward: I guess.
The Undiscovered Country
I never expected to know, say, the exact geometric average stock returns for the next few decades. What I had hoped to learn was the likely distribution of stock returns, e.g. an understanding of what are the most probable outcomes. With such a distribution I could create a retirement plan geared toward those most probable outcomes.
But near as I can tell, nobody has a clue as to what the distributions actually are. To define a distribution one has to understand the list of possible outcomes and be able to calculate how likely each is. Alas, no one knows how to do that for future asset behavior. So instead lots of folks spend their time trying to read the past.
The more sophisticated attempts to use the past to predict the future usually start with an economist who creates a model of how they believe the market works. They then feed in some set of 'starting state' information into this model and then the model makes predictions of future behavior. The model is tested by seeing if it can successfully predict behavior in past markets.
The fact that a model fits historical return data may or may not be a useful piece of information. For one thing I have no doubt that in many cases the models are tortured until they output the expected results. Even when models are tested against data sets that weren't used in their original creation the declaration of a 'match' is often astonishingly generous. Besides, is the future really going to be the same as the past? Can we be sure that factors that may have been critical in forming past market behaviors will continue to be as relevant in the future?
But even if someone does manage to come up with a model that accurately foretells the probable distribution of returns for some non-trivial period of time into the feature, how do I pick that particular model from its enumerable competitors? How do I test that it is telling me the truth? After all, the number of 'outcomes' in my case, say 30 years to retirement, are tiny. Just about any distribution will likely cover them, but that doesn't mean that the distribution was right or useful. So even if someone did have the 'answer', there's no way for me to know it which means that for me, the answer is as good as non-existent.
This Won't Hurt a Bit
So I have convinced myself that trying to come up with the 'right' distribution of stock returns or stock/bond correlations or any of the other numerous but critical inputs I need in order to put together a reasonable retirement strategy is a waste of time. But I still need estimates for these values otherwise I can't calculate how much money to save or where to save it.
So, lacking any other option, I have guessed. Sure, I have read a bunch of academic articles and looked at various estimates but in the end I recognized that all I am doing is guessing. So, for example, when I use the simulator I'll describe in a future article I recognize that the recommendations it makes for how much money we should save for retirement are nothing more than guesses on top of suppositions with a side of conjecture. I also recognize that because all of the inputs to the simulator are really just guesses it doesn't make sense to spend too much time agonizing about exactly what input to use for the simulator. It's not that one guess is as good as another but rather that at some point trying to improve a guess is just wasted effort.
For someone brought up in a world that believes science conquers all this is all terribly distressing. But, oh well, as one of my favorite philosophers once said "…always look on the bright side of life…"
[SWN] Bernstein, William J. "S.W.I.N.E.", Efficient Frontier, Summer 2002, http://www.efficientfrontier.com/ef/702/swine.htm.